SQuTR for MTEB

MTEB-ready packaging of SQuTR, a bilingual spoken-query-to-text retrieval benchmark.

It contains FiQA, HotpotQA, NQ, MedicalRetrieval, DuRetrieval, and T2Retrieval. Each collection includes clean, 20 dB, 10 dB, and 0 dB audio queries.

Query transcripts are kept for provenance but are not exposed to evaluated MTEB models.

SQuTR's generated data is CC BY-SA 4.0. The packaged corpora and qrels keep their original dataset licenses, including NQ's CC BY-NC-SA 3.0 terms. See the source dataset for details.

Source revision: 2f1b041e2e98e0d28ed68fbcf22126ef247eb719

Citation

@misc{li2026squtrrobustnessbenchmarkspoken,
  author = {Yuejie Li and Ke Yang and Yueying Hua and Berlin Chen and Jianhao Nie and Yueping He and Caixin Kang},
  title = {SQuTR: A Robustness Benchmark for Spoken Query to Text Retrieval under Acoustic Noise},
  year = {2026},
  eprint = {2602.12783},
  archivePrefix = {arXiv},
  primaryClass = {cs.IR},
  url = {https://arxiv.org/abs/2602.12783}
}